Claude Alt ⚡

A self-hosted, Claude.ai-inspired AI agent built on free LLM APIs — no subscriptions required. Runs on Railway (or any Docker host) and automatically falls back across 9+ providers when rate limits hit.

Live: claude-alt-production.up.railway.app
Org: The-No-Hands-company

---

What it does

• Chat with a capable AI agent backed by Groq, Gemini, Mistral, Cerebras, OpenRouter, Cohere, GitHub Models, Grok and more
• Agent can search the web, read/write/commit code, clone GitHub repos, run commands, generate images and more — all via a tool-calling loop
• Auto-fallback: hit a rate limit on one provider → silently moves to the next
• Smart routing: complex coding tasks get routed to the strongest available model; simple questions go to the fastest/cheapest
• Fully installable as a PWA on mobile and desktop

---

Stack

Layer	Tech
Backend	Python 3.12, FastAPI, Uvicorn
Agent loop	Custom streaming SSE agent with tool-call loop
Deployment	Docker, Railway
Frontend	Vanilla JS, Marked.js, Highlight.js

---

Quickstart (Railway)

1. Fork this repo
2. Create a new Railway project → Deploy from GitHub repo
3. Add environment variables (see below)
4. Deploy — app starts on the assigned Railway domain

Environment Variables

Variable	Required	Description
PROVIDER	Yes	auto (recommended) or a specific provider id
GROQ_API_KEY	Recommended	console.groq.com/keys — 14,400 req/day free
GEMINI_API_KEY	Recommended	aistudio.google.com/apikey — 1,500 req/day free
CEREBRAS_API_KEY	Optional	cloud.cerebras.ai
MISTRAL_API_KEY	Optional	console.mistral.ai
OPENROUTER_API_KEY	Optional	openrouter.ai/settings/keys
COHERE_API_KEY	Optional	dashboard.cohere.com
GITHUB_MODELS_TOKEN	Optional	github.com/settings/tokens — fine-grained, Models: read
NVIDIA_API_KEY	Optional	build.nvidia.com
GROK_API_KEY	Optional	console.x.ai (paid)
CLAUDE_API_KEY	Optional	console.anthropic.com (paid)
GH_TOKEN	Optional	GitHub PAT for default repo push access
GITHUB_REPO	Optional	Default repo URL for git operations
RATE_LIMIT_COOLDOWN	Optional	Seconds to cool a rate-limited provider (default: 60)

LLM7.io requires no key and is always available as a last-resort fallback.

---

Providers & Fallback Chain

When PROVIDER=auto, providers are tried in this order (first available wins):

1. LLM7.io        — keyless, always available (DeepSeek R1 7B)
2. Groq           — llama-3.3-70b-versatile,  14,400 RPD free
3. Cerebras       — llama-3.3-70b,             30 RPM free
4. Google Gemini  — gemini-2.0-flash,          1,500 RPD free
5. Mistral AI     — mistral-small-latest,       1B tok/mo free
6. OpenRouter     — llama-3.3-70b:free,         20 RPM free
7. Cohere         — command-r-plus,             1K req/mo free
8. GitHub Models  — Llama-3.3-70B,             150 RPD free
9. Grok (xAI)     — grok-3                     (paid)
10. Claude        — claude-sonnet-4             (paid)

Smart routing: tasks scored as high / medium / low complexity. High-complexity tasks (new projects, architecture, full implementations) skip cheap providers and try Claude/Grok/Gemini first.

---

Agent Tools

Built-in (no API key, instant)

Tool	What it does
get_time	Current time in any timezone, with aliases (e.g. "Sweden")
calculate	Safe math eval with full math module
weather	Current weather via wttr.in (no key)
currency	Live exchange rates via open.er-api.com (no key)
convert	30+ unit conversions (length, weight, volume, temp, data)
regex	Test regex patterns with match highlighting
base64	Encode / decode
json_format	Pretty-print and validate JSON
generate_image	Image generation via Pollinations.ai/flux (no key)

File & repo tools

Tool	What it does
write_file	Write file to session workdir; HTML/SVG renders as artifact
read_file	Read file contents (up to 300 lines preview)
list_files	Glob file listing
delete_file	Delete a file
clone_repo	Clone any GitHub repo (public or private with token)
run_command	Sandboxed shell — 256MB RAM cap, 10s CPU, 60s timeout
commit_push	Stage, commit and push to the target repo

Reasoning tools

Tool	What it does
think	Internal reasoning step (shown as 💭 in UI)
plan	Announce a build plan before executing (shown as 📋 card)
clarify	Ask the user focused questions before a complex task
web_search	DuckDuckGo search, top 5 results
respond	Final answer (ends the loop)

---

Agent Personas

Switch between modes from the strip below the header:

Persona	Best for	Temp	Provider tier
🤖 Assistant	General questions, daily tasks	0.2	Medium
💻 Coder	Code, architecture, debugging	0.1	High (best model first)
🔬 Researcher	Facts, deep dives, citations	0.3	High
🎨 Creative	Writing, storytelling, image prompts	0.8	Medium

---

Features & Roadmap

See ROADMAP.md for a full breakdown of what's implemented, what's in progress, and what's planned.